Matrix completion

[mortonjt]

Need to pass in the SoftImpute module into Gneiss

We'll want

• [ ] Cross validation with hyperparameter tuning on the minval
• [ ] A really solid model statistic to optimize over (i.e. mse or loglikelihood).
• [ ] A new OLS subclass that can perform the ilr-transform on the fly, so that we can actually fit pseudocounts from the matrix completion